The automation of honey pollen visual sorting overcomes the limitations of the conventional procedure helping the specialist in this time-consuming task. In this work, a novel and comprehensive Ground Truth of almost 19,000 images (from optical microscopy) of the 16 most abundant types of grains/pollen particles present in citrus and rosemary honey from Spain was constructed. This task was assisted by a HoneyApp (also developed herein) for the labelling and annotation process. Subsequently, the effectiveness of different pre-existing automatic pollen recognizers based on convolutional neural networks (CNN) (VGG16, VGG19, InceptionV3, Xception, ResNet50, DenseNet201, MobileNetV2 and EfficientNetV2M) was tested together with a new network proposed in this paper (PolleNetV1). The extreme complexity of those pre-existing CNN and extensive use of millions of parameters makes this new proposal especially promising. Although with a slightly lower accuracy (average 96%) in determining the relative frequencies of different types of pollen grains/particles, it has considerable advantages such as simplicity and ability to be included in the future functionality to automate pollen recognition in honey. This is the first step to finally achieving an objective tool that allows the correct labelling of any types of pollen in honey, thus contributing to its transparency in the market.

The classification of honey pollen grains is performed in order to classify honey according to its botanical origin, which is of great importance in terms of marketing. This visual work is currently done by human specialists counting and classifying the pollen grains in microscopic images. This is a hard, time-consuming, and subject to observer variability task. Thus, automated methods are required to overcome the limitations of the conventional procedure. This paper deals with the automatic classification of honey pollens using five representative Neural Networks coming from the ImageNet Challenge: VGG16, VGG19, ResNet50, InceptionV3 and Xception. The ground truth is composed of 9983 samples of 16 different types of pollens corresponding to citrus and rosemary pollens and its companions. The best result was obtained with the InceptionV3 network, achieving an accuracy of 98.15%, that outperforms the results obtained in previous works.

1 Departamento de Estadística e Investigación Operativa Aplicadas y Calidad, Universitat Politècnica de València, Camino de Vera s/n, 46022, Valencia, Spain 2 Instituto de Automática e Informática Industrial, Universitat Politècnica de València, Camino de Vera s/n, 46022, Valencia, Spain 3 Centro de Agroingeniería, Instituto Valenciano de Investigaciones Agrarias (IVIA), Cra. Moncada-Náquera km 5, 46113, Moncada, Spain

This paper proposes a robust deformable adaptive 2D model, based on computer vision methods, that automatically fits the body (ventral silhouette) of Bluefin tuna while swimming. Our model (without human intervention) adjusts to fish shape and size, obtaining fish orientation, bending to fit their flexion motion and has proved robust enough to overcome possible segmentation inaccuracies. Once the model has been successfully fitted to the fish it can ensure that the detected object is a tuna and not parts of fish or other objects. Automatic requirements of the fishing industry like biometric measurement, specimen counting or catch biomass estimation could then be addressed using a stereoscopic system and meaningful information extracted from our model. We also introduce a fitting procedure based on a fitting parameter – Fitting Error Index (FEI) – which permits us to know the quality of the results. In the experiments our model has achieved very high success rates (up to 90%) discriminating individuals in highly complex images acquired for us in real conditions in the Mediterranean Sea. Conclusions and future improvements to the proposed model are also discussed.

Nowadays, the detection, localization, and quantification of different kinds of features in an RGB image (segmentation) is extremely helpful for, e.g., process monitoring or customer product acceptance. In this article, some of the most commonly used RGB image segmentation approaches are compared in an orange quality control case study. Analysis of variance and correspondence analysis are combined for determining their most relevant differences and highlighting their pros and cons.

This paper presents an application of visual quality control in orange post-harvesting comparing two different approaches. These approaches correspond to two very different methodologies released in the area of Computer Vision. The first approach is based on Multivariate Image Analysis (MIA) and was originally developed for the detection of defects in random color textures. It uses Principal Component Analysis and the T2 statistic to map the defective areas. The second approach is based on Graph Image Segmentation (GIS). It is an efficient segmentation algorithm that uses a graph-based representation of the image and a predicate to measure the evidence of boundaries between adjacent regions. While the MIA approach performs novelty detection on defects using a trained model of sound color textures, the GIS approach is strictly an unsupervised method with no training required on sound or defective areas. Both methods are compared through experimental work performed on a ground truth of 120 samples of citrus coming from four different cultivars. Although the GIS approach is faster and achieves better results in defect detection, the MIA method provides less false detections and does not need to use the hypothesis that the bigger area in samples always correspond to the non-damaged area.

The success of oral surgery is subject to accurate advanced planning. In order to properly plan for dental surgery or a suitable implant placement, it is necessary an accurate segmentation of the jaw tissues: the teeth, the cortical bone, the trabecular core and over all, the inferior alveolar nerve. This manuscript presents a new automatic method that is based on fuzzy connectedness object extraction and mathematical morphology processing. The method uses computed tomography data to extract different views of the jaw: a pseudo-orthopantomographic view to estimate the path of the nerve and cross-sectional views to segment the jaw tissues. The method has been tested in a groundtruth set consisting of more than 9000 cross-sections from 20 different patients and has been evaluated using four similarity indicators (the Jaccard index, Dice's coefficient, point-to-point and point-to-curve distances), achieving promising results in all of them (0.726 ± 0.031, 0.840 ± 0.019, 0.144 ± 0.023 mm and 0.163 ± 0.025 mm, respectively). The method has proven to be significantly automated and accurate, with errors around 5% (of the diameter of the nerve), and is easily integrable in current dental planning systems.

In this work we study how we can use a novel model of spatial saliency (visual attention) combined with image features to significantly accelerate a scene recognition application and, at the same time, preserve recognition performance. To do so, we use a mobile robotlike application where scene recognition is carried out through the use of image features to characterize the different scenarios, and the Nearest Neighbor rule to carry out the classification. SIFT and SURF are two recent and competitive alternatives to image local featuring that we compare through extensive experimental work. Results from the experiments show that SIFT features perform significantly better than SURF features achieving important reductions in the size of the database of prototypes without significant losses in recognition performance, and thus, accelerating scene recognition. Also, from the experiments it is concluded that SURF features are less distinctive when using very large databases of interest points, as it occurs in the present case. Visual attention is the process by which the Human Visual System (HVS) is able to select from a given scene regions of interest that contain salient information, and thus, reduce the amount of information to be processed (Treisman, 1980; Koch, 1985). In the last decade, several computational models biologically motivated have been released to implement visual attention in image and video processing (Itti, 2000; Garcia-Diaz, 2008). Visual attention has also been used to improve object recognition and scene analysis (Bonaiuto, 2005; Walther, 2005). In this chapter, we study the utility of using a novel model of spatial saliency to improve a scene recognition application by reducing the amount of prototypes needed to carry out the classification task. The application is based on mobile robot-like video sequences taken in indoor facilities formed by several rooms and halls. The aim is to recognize the different scenarios in order to provide the mobile robot system with general location data. The visual attention approach is a novel model of bottom-up saliency that uses local phase information of the input data where the statistic information of second order is deleted to achieve a Retinoptical map of saliency. The proposed approach joints computational mechanisms of the two hypotheses largely accepted in early vision: first, the efficient coding

We consider the task of scene recognition, in the context of a robot-like navigation application, using a visual attention model of bottom-up saliency, invariant local features and visual landmarks, and the Nearest Neighbor rule for classification. Experimental work shows that important reductions in the number of prototypes used by the NN classifier can be achieved using saliency maps. We also present a novel approach to extract visual landmarks that uses the model of bottom-up saliency to localize interest points, and color centiles plus local binary pattern histograms to collect local description of them. In the experiments, this later approach outperforms SIFT features by achieving similar recognition results but further reductions in the size of the database of prototypes, thus providing bigger savings in computational costs.

One of the main problems in the post-harvest processing of citrus is the detection of visual defects in order to classify the fruit depending on their appearance. Species and cultivars of citrus present a high rate of unpredictability in texture and colour that makes it difficult to develop a general, unsupervised method able of perform this task. In this paper we study the use of a general approach that was originally developed for the detection of defects in random colour textures. It is based on a Multivariate Image Analysis strategy and uses Principal Component Analysis to extract a reference eigenspace from a matrix built by unfolding colour and spatial data from samples of defect-free peel. Test images are also unfolded and projected onto the reference eigenspace and the result is a score matrix which is used to compute defective maps based on the T^2 statistic. In addition, a multiresolution scheme is introduced in the original method to speed up the process. Unlike the techniques commonly used for the detection of defects in fruits, this is an unsupervised method that only needs a few samples to be trained. It is also a simple approach that is suitable for real-time compliance. Experimental work was performed on 120 samples of oranges and mandarins from four different cultivars: Clemenules, Marisol, Fortune, and Valencia. The success ratio for the detection of individual defects was 91.5%, while the classification ratio of damaged/sound samples was 94.2%. These results show that the studied method can be suitable for the task of citrus inspection.

In this paper we introduce an algorithm aimed to create a background model with multimodal support, which associates a confidence value to the obtained model. Our algorithm creates the model based on a criterion of motion, pixel behavior and pixel similarity with the scenes background. This method uses only three frames to create a first model without restrictions on the frame content. The model is adapted over time to reflect new situations and illumination changes in the scene. One approach to detect corrupt model is also mentioned. The goal of confidence value is to quantify the quality of the model after a number of frames have been used to build it. Quantitative experimental results are obtained with a well-known benchmark and compared to a classical background modelling algorithm, showing the benefits of our approach.

In this paper we study how the use of a novel model of bottom-up saliency (visual attention), based on local energy and color, can significantly accelerate scene recognition and, at the same time, preserve the recognition performance. To do so, we use a mobile robot-like application where scene recognition is performed through the use of SIFT features to characterize the different scenarios, and the Nearest Neighbor rule to carry out the classification. Experimental work shows that important reductions in the size of the database of prototypes can be achieved (17.6% of the original size) without significant losses in recognition performance (from 98.5% to 96.1%), thus accelerating the classification task.

In this paper, we study the feasibility of SIFT features for the tasks of object recognition and tracking within the framework of the IVSEE system design. The IVSEE system is intended to imitate the early functionalities of the human visual system in enclosed environments. The goal of this system is to be able to; perform basic object recognition, determine object states and spatial interrelations, and all of this engaged with a purposive system behavior (e.g. object tracking). To implement this system, we turn to well-known and state-of-the-art techniques from the literature, and choose SIFT features for the stages of object extraction and recognition. We have performed (and present here) experimental work carried out to determine the adequacy of these features for the system goals. Results confirm SIFT features as a good implementation choice.

We present an innovative way to simultaneously perform feature extraction and classification for the quality-control issue of surface grading by applying two multivariate statistical projection methods: SIMCA and PLS-DA. These tools have been applied to compress the color texture data that describe the visual appearance of surfaces (soft color texture descriptors) and to directly perform classification using statistics and predictions from the projection models. Experiments have been carried out using an extensive ceramic images database (VxC TSG) comprised of 14 different models, 42 surface classes, and 960 pieces. A factorial experimental design evaluated all the combinations of several factors affecting the accuracy rate. These factors include the tile model, color representation scheme (CIE Lab, CIE Luv, and RGB), and compression/classification approach (SIMCA and PLS-DA). Moreover, a logistic regression model is fitted from the experiments to compute accuracy estimates and study the effect of the factors on the accuracy rate. Results show that PLS-DA performs better than SIMCA, achieving a mean accuracy rate of 98.95%. These results outperform those obtained in a previous work where the soft color texture descriptors in combination with the CIE Lab color space and the k-NN classifier achieved an accuracy rate of 97.36%. (C) 2008 SPIE and IS&T [DOI: 10.1117/1.2957886]

This paper presents a novel approach to the question of surface grading, the soft color texture descriptors method. This method is extracted from an extensive evaluation process of several factors based on the use of two well established statistical tools: experimental design and logistic regression. The utility of different combinations of factors is evaluated in regard to the problem of automatic classification of materials such as ceramic tiles that need to be grouped according to homogeneous visual appearance, that is, the surface grading application. The set of factors includes the number of neighbors in the k-NN classifier (several values of k parameter), color space representation schemes (CIE Lab, CIE Luv, RGB, and grayscale), and color texture features (mean, standard deviation, 2nd–5th histogram moments). A factorial experimental design is performed testing all combinations of the above factors on a large image database of ceramic tiles. Accuracy estimates are computed using logistic regression to determine the best combinations of factors. From the point of view of machine learning the overall process conforms a wrapper approach able to select significant design choices (k parameter in k-NN classifier and color space) and carry out a feature selection within the set of color texture features at the same time. Experiments were repeated with alternate color texture schemes from the literature: color histograms and centile-LBP. Comparisons of methods are presented describing both accuracy estimates and runtimes.

In this paper we present an innovative way to simultaneously perform feature extraction and classification for the quality control issue of surface grading by applying two well known multivariate statistical projection tools (SIMCA and PLS-DA). These tools have been applied to compress the color texture data describing the visual appearance of surfaces (soft color texture descriptors) and to directly perform classification using statistics and predictions computed from the extracted projection models. Experiments have been carried out using an extensive image database of ceramic tiles (VxC TSG). This image database is comprised of 14 different models, 42 surface classes and 960 pieces. A factorial experimental design has been carried out to evaluate all the combinations of several factors affecting the accuracy rate. Factors include tile model, color representation scheme (CIE Lab, CIE Luv and RGB) and compression/classification approach (SIMCA and PLS-DA). In addition, a logistic regression model is fitted from the experiments to compute accuracy estimates and study the factors effect. The results show that PLS-DA performs better than SIMCA, achieving a mean accuracy rate of 98.95%. These results outperform those obtained in a previous work where the soft color texture descriptors in combination with the CIE Lab color space and the k-NN classi.er achieved a 97.36% of accuracy.

In this paper we present a new approach for the detection of defects in random colour textures. This approach is based on the use of the T2 statistic and it is derived from the MIA strategy (Multivariate Image Analysis) developed in recent years in the field of applied statistics. PCA analysis is used to extract a reference eigenspace from a matrix built by unfolding the RGB raw data of defect-free images. The unfolding is performed compiling colour and spatial information of pixels. New testing images are also unfolded and projected onto the reference eigenspace obtaining a score matrix used to compute the T2 images. These images are converted into defect maps which allow the location of defective pixels. Only very few samples are needed to perform unsupervised training. With regard to literature, the method uses one of the simplest approaches providing low computational costs.

This paper presents a new approach to the question of surface grading based on soft colour-texture descriptors and well known classifiers. These descriptors come from global image statistics computed in perceptually uniform colour spaces (CIE Lab or CIE Luv). The method has been extracted and validated using a statistical procedure based on experimental design and logistic regression. The method is not a new theoretical contribution, but we have found and demonstrate that a simple set of global statistics softly describing colour and texture properties, together with well-known classifiers, are powerful enough to meet stringent factory requirements for real-time and performance. These requirements are on-line inspection capability and 95% surface grading accuracy. The approach is also compared with two other methods in the surface grading literature; colour histograms [1] and centile-LBP [8]. This paper is an extension and in-depth development of ideas reported in a previous work [11].

In this paper we approach the problem of fast surface grading of flat pieces decorated with random patterns. The proposed method is based on the use of global statistics of color computed in the CIE Lab space. Two other fast methods based on color histograms [1] and Centile-LBP features [8] are introduced for comparison purposes. We used CIE Lab in order to provide accuracy and perceptual approach in color difference computation. Experiments with RGB were also carried out to study CIE Lab reliability. The ground truth was provided through an image database of ceramic tiles. Nevertheless, the approach is suitable to be extended to other random decorated surfaces like marble, granite, wood or textile stuff. The experiments make us to conclude that a simple collection of global statistics of color in the CIE Lab space is powerful enough to well discriminate surface grades. The average success surpasses 95% in most of the tests, improving literature methods and achieving factory compliance.